Interactive Relational Reinforcement Learning of concept semantics (extended abstract)

Abstract

We propose a novel approach to the machine learning of formal word sense, learned in interaction with human users using a new form of Relational Reinforcement Learning. The envisaged main application area of our framework is humanmachine communication, where a software agent or robot needs to understand concepts used by human users (e.g., in Natural Language Processing, HC1 or Information Retrieval). In contrast to traditional approaches to the machine learning and disambiguation of word meaning, our framework focuses on the interactive learning of concepts in a dialogue with the user and on the integration of rich formal background knowledge and dynamically adapted policy constraints in the learning process, which makes our approach suitable for dynamic interaction environments with varying word usage contexts. Keywords: Concept Learning, Statistical Relational Learning, Planning, Word-Sense Disambiguation, Dialogue Systems